Reinforcement and deep learning-based optimization of pose estimation techniques in single and multi-agent systems : application to aerial and mobile robots

Abstract

In this work, we will focus on optimizing pose estimation techniques for aerial and mobile robots in both single-agent and multi-agent systems. Novel approaches based on Deep Learning and Reinforcement Learning will be proposed to enhance accuracy and robustness. The thesis includes a comprehensive literature review, introducing software tools used in the research. Two approaches for single agent pose estimation, will be presented : the QR estimator for an adaptive version of the Extended Kalman Filter and the KalmanNet approach for a direct estimation of the Filter gain. The effectiveness of these approaches will be demonstrated through simulations. The investigation will then be extended to collaborative pose estimation in multi-agent systems. A novel approach will be also proposed which aims to improve accuracy and robustness by leveraging information from neighboring agents. The findings will be validated in a real-world simulation environment using ROS and Gazebo.